In validation

Market Oracle

A trading research system that was honest enough to tell me it didn't work

Most retail trading tools sell confidence. They surface signals, draw arrows, and imply an edge that rarely survives contact with real markets. The harder, more valuable question is rarely asked: does this strategy actually have an edge, or does it just look like one?

Market Oracle was built to answer that question before risking a dollar. Rather than assume the signals worked, the system was put through a 64-year backtest on survivorship-bias-free data — the kind of test that kills strategies that only ever looked good in hindsight.

  • Market data ingestion and a layered signal engine (mean-reversion, trend, and confirmation signals)
  • A backtesting pipeline run against decades of historical data, deliberately corrected for survivorship bias so dead companies still count
  • Per-signal edge measurement, so each component is judged on its own merits rather than hidden inside a blended score

The test was decisive: most of the technical signals showed no tradeable edge. Two mean-reversion signals held a small but persistent positive edge; the rest were flat to negative. It also exposed a structural flaw in the original logic — a buy gate that was AND-ing contradictory signals together, guaranteeing it rarely fired correctly.

Built withNode.js, Claude for the analysis layer, Alpaca for market data, QuantConnect for survivorship-free backtesting.

Anyone can build a tool that confirms what they hoped. This one was built to be falsifiable, and when the data said the edge wasn't there, that finding was the deliverable. That's the same rigor I bring to every system: test the assumption before you build on top of it.